Re: [PATCH] perf report: auto-detect branch stack sampling mode

[David Ahern <dsahern@gmail.com>]

On 2/24/12 8:28 AM, Stephane Eranian wrote:
> On Fri, Feb 24, 2012 at 4:24 PM, David Ahern<dsahern@gmail.com>  wrote:
>> On 2/24/12 2:40 AM, Stephane Eranian wrote:
>>>
>>>
>>> This patch adds auto-detection of samples with taken branch stacks.
>>> The auto-detection avoids having to specify the -b or --branch-stack
>>> option on the cmdline.
>>>
>>> The patch adds a new feature bit HEADER_BRANCH_STACK to mark the
>>> presence of branch stacks in samples.
>>>
>>> You can now do:
>>> $ perf record -b any noploop 2
>>> $ perf report
>>> # Events: 8K cycles
>>> #
>>> # Overhead  Command  Source Shared Object        Source Symbol  Target
>>> Shared Object       Target Symbol
>>> # ........  .......  ....................  ...................
>>>   ....................  ..................
>>> #
>>>      91.56%  noploop  noploop               [.] noploop
>>>    noploop  [.] noploop
>>>       0.42%  noploop  [kernel.kallsyms]     [k] __lock_acquire
>>>   [kernel.kallsyms]  [k] __lock_acquire
>>>
>>>
>>> To force regular reporting based on the instruction address:
>>> $ perf report --no-branch-stack
>>> #
>>> # Events: 2K cycles
>>> #
>>> # Overhead  Command      Shared Object                           Symbol
>>> # ........  .......  .................  ...............................
>>> #
>>>      92.03%  noploop  noploop            [.] noploop
>>>       1.00%  noploop  [kernel.kallsyms]  [k] lock_acquire
>>>
>>>
>>> Signed-off-by: Stephane Eranian<eranian@google.com>
>>> ---
>>>
>>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>>> index 1c49d4e..5e833a2 100644
>>> --- a/tools/perf/builtin-record.c
>>> +++ b/tools/perf/builtin-record.c
>>> @@ -473,6 +473,9 @@ static int __cmd_record(struct perf_record *rec, int
>>> argc, const char **argv)
>>>         if (!have_tracepoints(&evsel_list->entries))
>>>                 perf_header__clear_feat(&session->header,
>>> HEADER_TRACE_INFO);
>>>
>>> +       if (!rec->opts.branch_stack)
>>> +               perf_header__clear_feat(&session->header,
>>> HEADER_BRANCH_STACK);
>>
>>
>> branch tracing is user requested on, so shouldn't feature default off and
>> only be enabled when requested?
>>
> Well, what Ingo was suggesting is that perf report auto-detects whether or
> not branch mode is necessary by looking at the perf.data file. Most likely
> if you've recorded with -b, you are interested in a branch mode view rather
> that the instruction view (default). So all this does is elimintate the need
> to pass -b to perf report to enable branch mode.

Right. My point is that HEADER_BRANCH_STACK feature should be off by 
default and only enabled by builtin-record when branch stack is 
requested. You have the reverse -- enabled by default and off in record 
if not requested.

David